[DRBD-user] Still experiencing resource spikes

Lionel Sausin ls at numerigraphe.com
Tue Dec 18 11:55:25 CET 2012

Note: "permalinks" may not be as permanent as we would like,
direct links of old sources may well be a few messages off.

Le 18/12/2012 10:30, Andreas Kurz a écrit :
> That's quite a lot ... you also tuned vm.dirty_background_bytes and 
> vm.dirty_bytes to a reasonable low value? .... to avoid regular heavy 
> data write-out.... Regards, Andreas 
I'd look into this too, although with the given config and the default 
10% threshold for background writes of dirty pages, it should kick in 
for ~12 Gbytes of dirty. Let's say 120 Gbits get sent over the 20Gb/s 
link, that clean all up in a few seconds shouldn't it ?

I hope I'm not just adding noise, but if the kernel is old enough, there 
are cases where the write load is high enough for a sync operation to 
never end. Maybe that affects you?
In any case, it's worth running a "watch grep dirty /proc/vmstat" while 
the i/o block : if the cleaner is kicking in, nr_dirty should decrease 
down to almost zero. If it goes down an up again and again, maybe the 
write load is too high for a sync() to finish.

Lionel Sausin

More information about the drbd-user mailing list